90 likes | 226 Views
Quantile Curiosa. Gary Nan Tie Jan 1, 2010 gnantie@travelers.com. Introduction. The distribution function ( df ) of a random variable X is defined as: F X ( x ) = Pr[ X ≤ x ] The generalized inverse for a df , the quantile function q X (p ), is defined as:
E N D
Quantile Curiosa Gary Nan Tie Jan 1, 2010 gnantie@travelers.com
Introduction • The distribution function (df) of a random variable X is defined as: FX (x) = Pr[ X ≤ x ] • The generalized inverse for a df, the quantile function qX(p), is defined as: FX-1 (p) = inf { real x | FX (x) ≥ p } = sup { real x | FX (x) < p } • The ‘check’ function is defined as: ρq(r) = qr – r1{r < 0} for 0 ≤ q ≤ 1, e.g. ρ0.5 (r) = 0.5|r|
Curious folk result For an integrable rv X, the minimizer of E [ρq ( X - x) ] with respect to x, is the q-quantile of X. For an elementary proof, see [Hunter, Lange 1998, Appendix]
An even more curious result Suppose rv X has E[ |X| ] < ∞. Then the Fenchel-Legendre transform of the convex function Ψ(x) = E[ (x - X)+ ] is given by Ψ*(y) = sup xε Real ( xy - Ψ(x) ) = Integral from 0 to y of qX, if 0 ≤ y ≤ 1, and +∞ otherwise. Moreover, for 0 < y < 1, the supremum above is attained in x if and only if x is a y-quantile of X, that is x = qX(y). [Follmer, Schied, 2004, Lemma A.22]
Their connection! argminxεR{E [ρq( X - x) ] } = argminxεR{ qE[X] – qx – E[ (X – x)1{X-x < 0} ] } = argmaxxεR { qx + E[ (X – x)1{X-x < 0} ] } = argmaxxεR { xq - E[ (x - X)+ ] } = argmaxxεR { xq - Ψ(x) } = qX(q), for 0 < q < 1. So we have an elegant proof of the folk result via functional analysis!
Applications Quantile regression is a statistical technique used to estimate and make inference about conditional quantile functions [Koenker, Bassett, 1978]. Financial applications of quantile functions include asset pricing [Follmer, Schied, 2004], portfolio construction [Ma, Pohlman, 2005], [Bassett, Koenker, Kordas, 2004], risk management [Chernozhukov, 2002], [McNeil, Frey, Embrechts, 2005], and insurance [Denuit, Dhaene, Goovaerts, Kaas, 2005].
Some quantile function properties First order quantile ODE: dq/dp = 1/f(q) where q is the quantile function, 0 ≤ p ≤ 1, and f is the pdf. Second order non-linear ODE: d2q/dp2 = H(q) (dq/dp)2 where H(q) = -d/dqln{ f(q) }. For power series solutions, see [Steinbrecher, Shaw, 2007]. The quantile-characteristic function connection: φX(t) := E[ exp(itX) ] = Integral from 0 to 1 of exp(itqX) is explored via differentiation in [Shaw, McCabe, 2009]. May you discover more curious quantile properties!
References • Stochastic Finance 2nd Ed., Follmer and Schied, W de G 2004. • ‘An optimization transfer algorithm for quantile regression’, Hunter and Lange, 1998. • ‘Regression quantiles’, Koenker and Bassett, Econometrica, 46 33-50, 1978. • ‘Return forecasts and optimal portfolio construction’, Ma and Pohlman, 2005. • ‘Pessimistic portfolio allocation and Choquet expected utility’, Bassett, Koenker, and Kordas, 2004.
References (cont.) • ‘Extremalquantile regression’, Chernozhukov, 2002. • Quantitative Risk Management, McNeil, Frey and Embrechts, PUP 2005. • Actuarial Theory for Dependent Risks, Denuit, Dhaene, Goovaerts and Kaas, Wiley 2005. • ‘Quantile mechanics’, Steinbrecher and Shaw, 2007. • ‘Monte Carlo sampling given a characteristic function’, Shaw and McCabe, 2009.